1 //===- TargetTransformInfo.h ------------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This pass exposes codegen information to IR-level passes. Every
11 /// transformation that uses codegen information is broken into three parts:
12 /// 1. The IR-level analysis pass.
13 /// 2. The IR-level transformation interface which provides the needed
15 /// 3. Codegen-level implementation which uses target-specific hooks.
17 /// This file defines #2, which is the interface that IR-level transformations
18 /// use for querying the codegen.
20 //===----------------------------------------------------------------------===//
22 #ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
23 #define LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
25 #include "llvm/ADT/Optional.h"
26 #include "llvm/IR/Intrinsics.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/DataTypes.h"
36 class PreservedAnalyses;
41 /// \brief Information about a load/store intrinsic defined by the target.
42 struct MemIntrinsicInfo {
44 : ReadMem(false), WriteMem(false), Vol(false), MatchingId(0),
45 NumMemRefs(0), PtrVal(nullptr) {}
49 // Same Id is set by the target for corresponding load/store intrinsics.
50 unsigned short MatchingId;
55 /// \brief This pass provides access to the codegen interfaces that are needed
56 /// for IR-level transformations.
57 class TargetTransformInfo {
59 /// \brief Construct a TTI object using a type implementing the \c Concept
62 /// This is used by targets to construct a TTI wrapping their target-specific
63 /// implementaion that encodes appropriate costs for their target.
64 template <typename T> TargetTransformInfo(T Impl);
66 /// \brief Construct a baseline TTI object using a minimal implementation of
67 /// the \c Concept API below.
69 /// The TTI implementation will reflect the information in the DataLayout
70 /// provided if non-null.
71 explicit TargetTransformInfo(const DataLayout *DL);
73 // Provide move semantics.
74 TargetTransformInfo(TargetTransformInfo &&Arg);
75 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
77 // We need to define the destructor out-of-line to define our sub-classes
79 ~TargetTransformInfo();
81 /// \brief Handle the invalidation of this information.
83 /// When used as a result of \c TargetIRAnalysis this method will be called
84 /// when the function this was computed for changes. When it returns false,
85 /// the information is preserved across those changes.
86 bool invalidate(Function &, const PreservedAnalyses &) {
87 // FIXME: We should probably in some way ensure that the subtarget
88 // information for a function hasn't changed.
92 /// \name Generic Target Information
95 /// \brief Underlying constants for 'cost' values in this interface.
97 /// Many APIs in this interface return a cost. This enum defines the
98 /// fundamental values that should be used to interpret (and produce) those
99 /// costs. The costs are returned as an unsigned rather than a member of this
100 /// enumeration because it is expected that the cost of one IR instruction
101 /// may have a multiplicative factor to it or otherwise won't fit directly
102 /// into the enum. Moreover, it is common to sum or average costs which works
103 /// better as simple integral values. Thus this enum only provides constants.
105 /// Note that these costs should usually reflect the intersection of code-size
106 /// cost and execution cost. A free instruction is typically one that folds
107 /// into another instruction. For example, reg-to-reg moves can often be
108 /// skipped by renaming the registers in the CPU, but they still are encoded
109 /// and thus wouldn't be considered 'free' here.
110 enum TargetCostConstants {
111 TCC_Free = 0, ///< Expected to fold away in lowering.
112 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
113 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
116 /// \brief Estimate the cost of a specific operation when lowered.
118 /// Note that this is designed to work on an arbitrary synthetic opcode, and
119 /// thus work for hypothetical queries before an instruction has even been
120 /// formed. However, this does *not* work for GEPs, and must not be called
121 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
122 /// analyzing a GEP's cost required more information.
124 /// Typically only the result type is required, and the operand type can be
125 /// omitted. However, if the opcode is one of the cast instructions, the
126 /// operand type is required.
128 /// The returned cost is defined in terms of \c TargetCostConstants, see its
129 /// comments for a detailed explanation of the cost values.
130 unsigned getOperationCost(unsigned Opcode, Type *Ty,
131 Type *OpTy = nullptr) const;
133 /// \brief Estimate the cost of a GEP operation when lowered.
135 /// The contract for this function is the same as \c getOperationCost except
136 /// that it supports an interface that provides extra information specific to
137 /// the GEP operation.
138 unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) const;
140 /// \brief Estimate the cost of a function call when lowered.
142 /// The contract for this is the same as \c getOperationCost except that it
143 /// supports an interface that provides extra information specific to call
146 /// This is the most basic query for estimating call cost: it only knows the
147 /// function type and (potentially) the number of arguments at the call site.
148 /// The latter is only interesting for varargs function types.
149 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
151 /// \brief Estimate the cost of calling a specific function when lowered.
153 /// This overload adds the ability to reason about the particular function
154 /// being called in the event it is a library call with special lowering.
155 unsigned getCallCost(const Function *F, int NumArgs = -1) const;
157 /// \brief Estimate the cost of calling a specific function when lowered.
159 /// This overload allows specifying a set of candidate argument values.
160 unsigned getCallCost(const Function *F,
161 ArrayRef<const Value *> Arguments) const;
163 /// \brief Estimate the cost of an intrinsic when lowered.
165 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
166 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
167 ArrayRef<Type *> ParamTys) const;
169 /// \brief Estimate the cost of an intrinsic when lowered.
171 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
172 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
173 ArrayRef<const Value *> Arguments) const;
175 /// \brief Estimate the cost of a given IR user when lowered.
177 /// This can estimate the cost of either a ConstantExpr or Instruction when
178 /// lowered. It has two primary advantages over the \c getOperationCost and
179 /// \c getGEPCost above, and one significant disadvantage: it can only be
180 /// used when the IR construct has already been formed.
182 /// The advantages are that it can inspect the SSA use graph to reason more
183 /// accurately about the cost. For example, all-constant-GEPs can often be
184 /// folded into a load or other instruction, but if they are used in some
185 /// other context they may not be folded. This routine can distinguish such
188 /// The returned cost is defined in terms of \c TargetCostConstants, see its
189 /// comments for a detailed explanation of the cost values.
190 unsigned getUserCost(const User *U) const;
192 /// \brief hasBranchDivergence - Return true if branch divergence exists.
193 /// Branch divergence has a significantly negative impact on GPU performance
194 /// when threads in the same wavefront take different paths due to conditional
196 bool hasBranchDivergence() const;
198 /// \brief Test whether calls to a function lower to actual program function
201 /// The idea is to test whether the program is likely to require a 'call'
202 /// instruction or equivalent in order to call the given function.
204 /// FIXME: It's not clear that this is a good or useful query API. Client's
205 /// should probably move to simpler cost metrics using the above.
206 /// Alternatively, we could split the cost interface into distinct code-size
207 /// and execution-speed costs. This would allow modelling the core of this
208 /// query more accurately as a call is a single small instruction, but
209 /// incurs significant execution cost.
210 bool isLoweredToCall(const Function *F) const;
212 /// Parameters that control the generic loop unrolling transformation.
213 struct UnrollingPreferences {
214 /// The cost threshold for the unrolled loop, compared to
215 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
216 /// The unrolling factor is set such that the unrolled loop body does not
217 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
220 /// The cost threshold for the unrolled loop when optimizing for size (set
221 /// to UINT_MAX to disable).
222 unsigned OptSizeThreshold;
223 /// The cost threshold for the unrolled loop, like Threshold, but used
224 /// for partial/runtime unrolling (set to UINT_MAX to disable).
225 unsigned PartialThreshold;
226 /// The cost threshold for the unrolled loop when optimizing for size, like
227 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
228 /// UINT_MAX to disable).
229 unsigned PartialOptSizeThreshold;
230 /// A forced unrolling factor (the number of concatenated bodies of the
231 /// original loop in the unrolled loop body). When set to 0, the unrolling
232 /// transformation will select an unrolling factor based on the current cost
233 /// threshold and other factors.
235 // Set the maximum unrolling factor. The unrolling factor may be selected
236 // using the appropriate cost threshold, but may not exceed this number
237 // (set to UINT_MAX to disable). This does not apply in cases where the
238 // loop is being fully unrolled.
240 /// Allow partial unrolling (unrolling of loops to expand the size of the
241 /// loop body, not only to eliminate small constant-trip-count loops).
243 /// Allow runtime unrolling (unrolling of loops to expand the size of the
244 /// loop body even when the number of loop iterations is not known at
249 /// \brief Get target-customized preferences for the generic loop unrolling
250 /// transformation. The caller will initialize UP with the current
251 /// target-independent defaults.
252 void getUnrollingPreferences(const Function *F, Loop *L,
253 UnrollingPreferences &UP) const;
257 /// \name Scalar Target Information
260 /// \brief Flags indicating the kind of support for population count.
262 /// Compared to the SW implementation, HW support is supposed to
263 /// significantly boost the performance when the population is dense, and it
264 /// may or may not degrade performance if the population is sparse. A HW
265 /// support is considered as "Fast" if it can outperform, or is on a par
266 /// with, SW implementation when the population is sparse; otherwise, it is
267 /// considered as "Slow".
268 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
270 /// \brief Return true if the specified immediate is legal add immediate, that
271 /// is the target has add instructions which can add a register with the
272 /// immediate without having to materialize the immediate into a register.
273 bool isLegalAddImmediate(int64_t Imm) const;
275 /// \brief Return true if the specified immediate is legal icmp immediate,
276 /// that is the target has icmp instructions which can compare a register
277 /// against the immediate without having to materialize the immediate into a
279 bool isLegalICmpImmediate(int64_t Imm) const;
281 /// \brief Return true if the addressing mode represented by AM is legal for
282 /// this target, for a load/store of the specified type.
283 /// The type may be VoidTy, in which case only return true if the addressing
284 /// mode is legal for a load/store of any legal type.
285 /// TODO: Handle pre/postinc as well.
286 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
287 bool HasBaseReg, int64_t Scale) const;
289 /// \brief Return true if the target works with masked instruction
290 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
291 /// AVX-512 architecture will also allow masks for non-consecutive memory
293 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
294 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
296 /// \brief Return the cost of the scaling factor used in the addressing
297 /// mode represented by AM for this target, for a load/store
298 /// of the specified type.
299 /// If the AM is supported, the return value must be >= 0.
300 /// If the AM is not supported, it returns a negative value.
301 /// TODO: Handle pre/postinc as well.
302 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
303 bool HasBaseReg, int64_t Scale) const;
305 /// \brief Return true if it's free to truncate a value of type Ty1 to type
306 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
307 /// by referencing its sub-register AX.
308 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
310 /// \brief Return true if this type is legal.
311 bool isTypeLegal(Type *Ty) const;
313 /// \brief Returns the target's jmp_buf alignment in bytes.
314 unsigned getJumpBufAlignment() const;
316 /// \brief Returns the target's jmp_buf size in bytes.
317 unsigned getJumpBufSize() const;
319 /// \brief Return true if switches should be turned into lookup tables for the
321 bool shouldBuildLookupTables() const;
323 /// \brief Return hardware support for population count.
324 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
326 /// \brief Return true if the hardware has a fast square-root instruction.
327 bool haveFastSqrt(Type *Ty) const;
329 /// \brief Return the expected cost of materializing for the given integer
330 /// immediate of the specified type.
331 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
333 /// \brief Return the expected cost of materialization for the given integer
334 /// immediate of the specified type for a given instruction. The cost can be
335 /// zero if the immediate can be folded into the specified instruction.
336 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
338 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
342 /// \name Vector Target Information
345 /// \brief The various kinds of shuffle patterns for vector queries.
347 SK_Broadcast, ///< Broadcast element 0 to all other elements.
348 SK_Reverse, ///< Reverse the order of the vector.
349 SK_Alternate, ///< Choose alternate elements from vector.
350 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
351 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
354 /// \brief Additional information about an operand's possible values.
355 enum OperandValueKind {
356 OK_AnyValue, // Operand can have any value.
357 OK_UniformValue, // Operand is uniform (splat of a value).
358 OK_UniformConstantValue, // Operand is uniform constant.
359 OK_NonUniformConstantValue // Operand is a non uniform constant value.
362 /// \brief Additional properties of an operand's values.
363 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
365 /// \return The number of scalar or vector registers that the target has.
366 /// If 'Vectors' is true, it returns the number of vector registers. If it is
367 /// set to false, it returns the number of scalar registers.
368 unsigned getNumberOfRegisters(bool Vector) const;
370 /// \return The width of the largest scalar or vector register type.
371 unsigned getRegisterBitWidth(bool Vector) const;
373 /// \return The maximum interleave factor that any transform should try to
374 /// perform for this target. This number depends on the level of parallelism
375 /// and the number of execution units in the CPU.
376 unsigned getMaxInterleaveFactor() const;
378 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
380 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
381 OperandValueKind Opd1Info = OK_AnyValue,
382 OperandValueKind Opd2Info = OK_AnyValue,
383 OperandValueProperties Opd1PropInfo = OP_None,
384 OperandValueProperties Opd2PropInfo = OP_None) const;
386 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
387 /// The index and subtype parameters are used by the subvector insertion and
388 /// extraction shuffle kinds.
389 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
390 Type *SubTp = nullptr) const;
392 /// \return The expected cost of cast instructions, such as bitcast, trunc,
394 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
396 /// \return The expected cost of control-flow related instructions such as
398 unsigned getCFInstrCost(unsigned Opcode) const;
400 /// \returns The expected cost of compare and select instructions.
401 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
402 Type *CondTy = nullptr) const;
404 /// \return The expected cost of vector Insert and Extract.
405 /// Use -1 to indicate that there is no information on the index value.
406 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
407 unsigned Index = -1) const;
409 /// \return The cost of Load and Store instructions.
410 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
411 unsigned AddressSpace) const;
413 /// \return The cost of masked Load and Store instructions.
414 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
415 unsigned AddressSpace) const;
417 /// \brief Calculate the cost of performing a vector reduction.
419 /// This is the cost of reducing the vector value of type \p Ty to a scalar
420 /// value using the operation denoted by \p Opcode. The form of the reduction
421 /// can either be a pairwise reduction or a reduction that splits the vector
422 /// at every reduction level.
426 /// ((v0+v1), (v2, v3), undef, undef)
429 /// ((v0+v2), (v1+v3), undef, undef)
430 unsigned getReductionCost(unsigned Opcode, Type *Ty,
431 bool IsPairwiseForm) const;
433 /// \returns The cost of Intrinsic instructions.
434 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
435 ArrayRef<Type *> Tys) const;
437 /// \returns The number of pieces into which the provided type must be
438 /// split during legalization. Zero is returned when the answer is unknown.
439 unsigned getNumberOfParts(Type *Tp) const;
441 /// \returns The cost of the address computation. For most targets this can be
442 /// merged into the instruction indexing mode. Some targets might want to
443 /// distinguish between address computation for memory operations on vector
444 /// types and scalar types. Such targets should override this function.
445 /// The 'IsComplex' parameter is a hint that the address computation is likely
446 /// to involve multiple instructions and as such unlikely to be merged into
447 /// the address indexing mode.
448 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
450 /// \returns The cost, if any, of keeping values of the given types alive
453 /// Some types may require the use of register classes that do not have
454 /// any callee-saved registers, so would require a spill and fill.
455 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
457 /// \returns True if the intrinsic is a supported memory intrinsic. Info
458 /// will contain additional information - whether the intrinsic may write
459 /// or read to memory, volatility and the pointer. Info is undefined
460 /// if false is returned.
461 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
463 /// \returns A value which is the result of the given memory intrinsic. New
464 /// instructions may be created to extract the result from the given intrinsic
465 /// memory operation. Returns nullptr if the target cannot create a result
466 /// from the given intrinsic.
467 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
468 Type *ExpectedType) const;
473 /// \brief The abstract base class used to type erase specific TTI
477 /// \brief The template model for the base class which wraps a concrete
478 /// implementation in a type erased interface.
479 template <typename T> class Model;
481 std::unique_ptr<Concept> TTIImpl;
484 class TargetTransformInfo::Concept {
486 virtual ~Concept() = 0;
488 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
489 virtual unsigned getGEPCost(const Value *Ptr,
490 ArrayRef<const Value *> Operands) = 0;
491 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
492 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
493 virtual unsigned getCallCost(const Function *F,
494 ArrayRef<const Value *> Arguments) = 0;
495 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
496 ArrayRef<Type *> ParamTys) = 0;
497 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
498 ArrayRef<const Value *> Arguments) = 0;
499 virtual unsigned getUserCost(const User *U) = 0;
500 virtual bool hasBranchDivergence() = 0;
501 virtual bool isLoweredToCall(const Function *F) = 0;
502 virtual void getUnrollingPreferences(const Function *F, Loop *L,
503 UnrollingPreferences &UP) = 0;
504 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
505 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
506 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
507 int64_t BaseOffset, bool HasBaseReg,
509 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
510 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
511 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
512 int64_t BaseOffset, bool HasBaseReg,
514 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
515 virtual bool isTypeLegal(Type *Ty) = 0;
516 virtual unsigned getJumpBufAlignment() = 0;
517 virtual unsigned getJumpBufSize() = 0;
518 virtual bool shouldBuildLookupTables() = 0;
519 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
520 virtual bool haveFastSqrt(Type *Ty) = 0;
521 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
522 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
524 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
525 const APInt &Imm, Type *Ty) = 0;
526 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
527 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
528 virtual unsigned getMaxInterleaveFactor() = 0;
530 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
531 OperandValueKind Opd2Info,
532 OperandValueProperties Opd1PropInfo,
533 OperandValueProperties Opd2PropInfo) = 0;
534 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
536 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
537 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
538 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
540 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
542 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
544 unsigned AddressSpace) = 0;
545 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
547 unsigned AddressSpace) = 0;
548 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
549 bool IsPairwiseForm) = 0;
550 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
551 ArrayRef<Type *> Tys) = 0;
552 virtual unsigned getNumberOfParts(Type *Tp) = 0;
553 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
554 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
555 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
556 MemIntrinsicInfo &Info) = 0;
557 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
558 Type *ExpectedType) = 0;
561 template <typename T>
562 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
566 Model(T Impl) : Impl(std::move(Impl)) {}
569 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
570 return Impl.getOperationCost(Opcode, Ty, OpTy);
572 unsigned getGEPCost(const Value *Ptr,
573 ArrayRef<const Value *> Operands) override {
574 return Impl.getGEPCost(Ptr, Operands);
576 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
577 return Impl.getCallCost(FTy, NumArgs);
579 unsigned getCallCost(const Function *F, int NumArgs) override {
580 return Impl.getCallCost(F, NumArgs);
582 unsigned getCallCost(const Function *F,
583 ArrayRef<const Value *> Arguments) override {
584 return Impl.getCallCost(F, Arguments);
586 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
587 ArrayRef<Type *> ParamTys) override {
588 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
590 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
591 ArrayRef<const Value *> Arguments) override {
592 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
594 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
595 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
596 bool isLoweredToCall(const Function *F) override {
597 return Impl.isLoweredToCall(F);
599 void getUnrollingPreferences(const Function *F, Loop *L,
600 UnrollingPreferences &UP) override {
601 return Impl.getUnrollingPreferences(F, L, UP);
603 bool isLegalAddImmediate(int64_t Imm) override {
604 return Impl.isLegalAddImmediate(Imm);
606 bool isLegalICmpImmediate(int64_t Imm) override {
607 return Impl.isLegalICmpImmediate(Imm);
609 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
610 bool HasBaseReg, int64_t Scale) override {
611 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
614 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
615 return Impl.isLegalMaskedStore(DataType, Consecutive);
617 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
618 return Impl.isLegalMaskedLoad(DataType, Consecutive);
620 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
621 bool HasBaseReg, int64_t Scale) override {
622 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, Scale);
624 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
625 return Impl.isTruncateFree(Ty1, Ty2);
627 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
628 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
629 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
630 bool shouldBuildLookupTables() override {
631 return Impl.shouldBuildLookupTables();
633 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
634 return Impl.getPopcntSupport(IntTyWidthInBit);
636 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
637 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
638 return Impl.getIntImmCost(Imm, Ty);
640 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
642 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
644 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
646 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
648 unsigned getNumberOfRegisters(bool Vector) override {
649 return Impl.getNumberOfRegisters(Vector);
651 unsigned getRegisterBitWidth(bool Vector) override {
652 return Impl.getRegisterBitWidth(Vector);
654 unsigned getMaxInterleaveFactor() override {
655 return Impl.getMaxInterleaveFactor();
658 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
659 OperandValueKind Opd2Info,
660 OperandValueProperties Opd1PropInfo,
661 OperandValueProperties Opd2PropInfo) override {
662 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
663 Opd1PropInfo, Opd2PropInfo);
665 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
666 Type *SubTp) override {
667 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
669 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
670 return Impl.getCastInstrCost(Opcode, Dst, Src);
672 unsigned getCFInstrCost(unsigned Opcode) override {
673 return Impl.getCFInstrCost(Opcode);
675 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
676 Type *CondTy) override {
677 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
679 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
680 unsigned Index) override {
681 return Impl.getVectorInstrCost(Opcode, Val, Index);
683 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
684 unsigned AddressSpace) override {
685 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
687 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
688 unsigned AddressSpace) override {
689 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
691 unsigned getReductionCost(unsigned Opcode, Type *Ty,
692 bool IsPairwiseForm) override {
693 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
695 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
696 ArrayRef<Type *> Tys) override {
697 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
699 unsigned getNumberOfParts(Type *Tp) override {
700 return Impl.getNumberOfParts(Tp);
702 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
703 return Impl.getAddressComputationCost(Ty, IsComplex);
705 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
706 return Impl.getCostOfKeepingLiveOverCall(Tys);
708 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
709 MemIntrinsicInfo &Info) override {
710 return Impl.getTgtMemIntrinsic(Inst, Info);
712 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
713 Type *ExpectedType) override {
714 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
718 template <typename T>
719 TargetTransformInfo::TargetTransformInfo(T Impl)
720 : TTIImpl(new Model<T>(Impl)) {}
722 /// \brief Analysis pass providing the \c TargetTransformInfo.
724 /// The core idea of the TargetIRAnalysis is to expose an interface through
725 /// which LLVM targets can analyze and provide information about the middle
726 /// end's target-independent IR. This supports use cases such as target-aware
727 /// cost modeling of IR constructs.
729 /// This is a function analysis because much of the cost modeling for targets
730 /// is done in a subtarget specific way and LLVM supports compiling different
731 /// functions targeting different subtargets in order to support runtime
732 /// dispatch according to the observed subtarget.
733 class TargetIRAnalysis {
735 typedef TargetTransformInfo Result;
737 /// \brief Opaque, unique identifier for this analysis pass.
738 static void *ID() { return (void *)&PassID; }
740 /// \brief Provide access to a name for this pass for debugging purposes.
741 static StringRef name() { return "TargetIRAnalysis"; }
743 /// \brief Default construct a target IR analysis.
745 /// This will use the module's datalayout to construct a baseline
746 /// conservative TTI result.
749 /// \brief Construct an IR analysis pass around a target-provide callback.
751 /// The callback will be called with a particular function for which the TTI
752 /// is needed and must return a TTI object for that function.
753 TargetIRAnalysis(std::function<Result(Function &)> TTICallback);
755 // Value semantics. We spell out the constructors for MSVC.
756 TargetIRAnalysis(const TargetIRAnalysis &Arg)
757 : TTICallback(Arg.TTICallback) {}
758 TargetIRAnalysis(TargetIRAnalysis &&Arg)
759 : TTICallback(std::move(Arg.TTICallback)) {}
760 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
761 TTICallback = RHS.TTICallback;
764 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
765 TTICallback = std::move(RHS.TTICallback);
769 Result run(Function &F);
774 /// \brief The callback used to produce a result.
776 /// We use a completely opaque callback so that targets can provide whatever
777 /// mechanism they desire for constructing the TTI for a given function.
779 /// FIXME: Should we really use std::function? It's relatively inefficient.
780 /// It might be possible to arrange for even stateful callbacks to outlive
781 /// the analysis and thus use a function_ref which would be lighter weight.
782 /// This may also be less error prone as the callback is likely to reference
783 /// the external TargetMachine, and that reference needs to never dangle.
784 std::function<Result(Function &)> TTICallback;
786 /// \brief Helper function used as the callback in the default constructor.
787 static Result getDefaultTTI(Function &F);
790 /// \brief Wrapper pass for TargetTransformInfo.
792 /// This pass can be constructed from a TTI object which it stores internally
793 /// and is queried by passes.
794 class TargetTransformInfoWrapperPass : public ImmutablePass {
795 TargetIRAnalysis TIRA;
796 Optional<TargetTransformInfo> TTI;
798 virtual void anchor();
803 /// \brief We must provide a default constructor for the pass but it should
806 /// Use the constructor below or call one of the creation routines.
807 TargetTransformInfoWrapperPass();
809 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
811 TargetTransformInfo &getTTI(Function &F);
814 /// \brief Create an analysis pass wrapper around a TTI object.
816 /// This analysis pass just holds the TTI instance and makes it available to
818 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
820 } // End llvm namespace